1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx2-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 const int8_t* i9 = input[9];
79 assert(i9 != NULL);
80 if XNN_UNPREDICTABLE(i9 != zero) {
81 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
82 }
83 const int8_t* i10 = input[10];
84 assert(i10 != NULL);
85 if XNN_UNPREDICTABLE(i10 != zero) {
86 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
87 }
88 const int8_t* i11 = input[11];
89 assert(i11 != NULL);
90 if XNN_UNPREDICTABLE(i11 != zero) {
91 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
92 }
93 const int8_t* i12 = input[12];
94 assert(i12 != NULL);
95 if XNN_UNPREDICTABLE(i12 != zero) {
96 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
97 }
98 const int8_t* i13 = input[13];
99 assert(i13 != NULL);
100 if XNN_UNPREDICTABLE(i13 != zero) {
101 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
102 }
103 const int8_t* i14 = input[14];
104 assert(i14 != NULL);
105 if XNN_UNPREDICTABLE(i14 != zero) {
106 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
107 }
108 const int8_t* i15 = input[15];
109 assert(i15 != NULL);
110 if XNN_UNPREDICTABLE(i15 != zero) {
111 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
112 }
113 const int8_t* i16 = input[16];
114 assert(i16 != NULL);
115 if XNN_UNPREDICTABLE(i16 != zero) {
116 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
117 }
118 const int8_t* i17 = input[17];
119 assert(i17 != NULL);
120 if XNN_UNPREDICTABLE(i17 != zero) {
121 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
122 }
123 const int8_t* i18 = input[18];
124 assert(i18 != NULL);
125 if XNN_UNPREDICTABLE(i18 != zero) {
126 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
127 }
128 const int8_t* i19 = input[19];
129 assert(i19 != NULL);
130 if XNN_UNPREDICTABLE(i19 != zero) {
131 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
132 }
133 const int8_t* i20 = input[20];
134 assert(i20 != NULL);
135 if XNN_UNPREDICTABLE(i20 != zero) {
136 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
137 }
138 const int8_t* i21 = input[21];
139 assert(i21 != NULL);
140 if XNN_UNPREDICTABLE(i21 != zero) {
141 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
142 }
143 const int8_t* i22 = input[22];
144 assert(i22 != NULL);
145 if XNN_UNPREDICTABLE(i22 != zero) {
146 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
147 }
148 const int8_t* i23 = input[23];
149 assert(i23 != NULL);
150 if XNN_UNPREDICTABLE(i23 != zero) {
151 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
152 }
153 const int8_t* i24 = input[24];
154 assert(i24 != NULL);
155 if XNN_UNPREDICTABLE(i24 != zero) {
156 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
157 }
158 input = (const int8_t**) ((uintptr_t) input + input_stride);
159
160 size_t c = channels;
161 const void* w = weights;
162 for (; c >= 32; c -= 32) {
163 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
164 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 8));
165 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 16));
166 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 24));
167
168
169 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
170 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
171 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
172 const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 8 * sizeof(int8_t))));
173 const __m256i vi0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 16)));
174 const __m256i vk0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
175 const __m256i vi0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 24)));
176 const __m256i vk0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 24 * sizeof(int8_t))));
177 i0 += 32;
178
179 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
180 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi0x89ABCDEF, vk0x89ABCDEF));
181 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi0xGHIJKLMN, vk0xGHIJKLMN));
182 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi0xOPQRSTUV, vk0xOPQRSTUV));
183
184 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
185 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
186 const __m256i vi1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
187 const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 40 * sizeof(int8_t))));
188 const __m256i vi1xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 16)));
189 const __m256i vk1xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
190 const __m256i vi1xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 24)));
191 const __m256i vk1xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 56 * sizeof(int8_t))));
192 i1 += 32;
193
194 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
195 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF));
196 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi1xGHIJKLMN, vk1xGHIJKLMN));
197 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi1xOPQRSTUV, vk1xOPQRSTUV));
198
199 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
200 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
201 const __m256i vi2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
202 const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 72 * sizeof(int8_t))));
203 const __m256i vi2xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 16)));
204 const __m256i vk2xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
205 const __m256i vi2xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 24)));
206 const __m256i vk2xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 88 * sizeof(int8_t))));
207 i2 += 32;
208
209 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
210 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF));
211 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi2xGHIJKLMN, vk2xGHIJKLMN));
212 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi2xOPQRSTUV, vk2xOPQRSTUV));
213
214 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
215 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
216 const __m256i vi3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
217 const __m256i vk3x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 104 * sizeof(int8_t))));
218 const __m256i vi3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 16)));
219 const __m256i vk3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
220 const __m256i vi3xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 24)));
221 const __m256i vk3xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 120 * sizeof(int8_t))));
222 i3 += 32;
223
224 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
225 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi3x89ABCDEF, vk3x89ABCDEF));
226 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN));
227 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi3xOPQRSTUV, vk3xOPQRSTUV));
228
229 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
230 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
231 const __m256i vi4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
232 const __m256i vk4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 136 * sizeof(int8_t))));
233 const __m256i vi4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 16)));
234 const __m256i vk4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
235 const __m256i vi4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 24)));
236 const __m256i vk4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 152 * sizeof(int8_t))));
237 i4 += 32;
238
239 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
240 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF));
241 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN));
242 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi4xOPQRSTUV, vk4xOPQRSTUV));
243
244 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
245 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
246 const __m256i vi5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
247 const __m256i vk5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 168 * sizeof(int8_t))));
248 const __m256i vi5xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 16)));
249 const __m256i vk5xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
250 const __m256i vi5xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 24)));
251 const __m256i vk5xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 184 * sizeof(int8_t))));
252 i5 += 32;
253
254 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
255 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF));
256 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi5xGHIJKLMN, vk5xGHIJKLMN));
257 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi5xOPQRSTUV, vk5xOPQRSTUV));
258
259 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
260 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
261 const __m256i vi6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
262 const __m256i vk6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 200 * sizeof(int8_t))));
263 const __m256i vi6xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 16)));
264 const __m256i vk6xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
265 const __m256i vi6xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 24)));
266 const __m256i vk6xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 216 * sizeof(int8_t))));
267 i6 += 32;
268
269 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
270 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF));
271 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi6xGHIJKLMN, vk6xGHIJKLMN));
272 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi6xOPQRSTUV, vk6xOPQRSTUV));
273
274 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
275 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
276 const __m256i vi7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 8)));
277 const __m256i vk7x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 232 * sizeof(int8_t))));
278 const __m256i vi7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 16)));
279 const __m256i vk7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
280 const __m256i vi7xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 24)));
281 const __m256i vk7xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 248 * sizeof(int8_t))));
282 i7 += 32;
283
284 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
285 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi7x89ABCDEF, vk7x89ABCDEF));
286 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN));
287 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi7xOPQRSTUV, vk7xOPQRSTUV));
288
289 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
290 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
291 const __m256i vi8x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 8)));
292 const __m256i vk8x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 264 * sizeof(int8_t))));
293 const __m256i vi8xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 16)));
294 const __m256i vk8xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
295 const __m256i vi8xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 24)));
296 const __m256i vk8xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 280 * sizeof(int8_t))));
297 i8 += 32;
298
299 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
300 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi8x89ABCDEF, vk8x89ABCDEF));
301 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi8xGHIJKLMN, vk8xGHIJKLMN));
302 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi8xOPQRSTUV, vk8xOPQRSTUV));
303
304 const __m256i vi9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i9));
305 const __m256i vk9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t))));
306 const __m256i vi9x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i9 + 8)));
307 const __m256i vk9x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 296 * sizeof(int8_t))));
308 const __m256i vi9xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i9 + 16)));
309 const __m256i vk9xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 304 * sizeof(int8_t))));
310 const __m256i vi9xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i9 + 24)));
311 const __m256i vk9xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 312 * sizeof(int8_t))));
312 i9 += 32;
313
314 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
315 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi9x89ABCDEF, vk9x89ABCDEF));
316 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi9xGHIJKLMN, vk9xGHIJKLMN));
317 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi9xOPQRSTUV, vk9xOPQRSTUV));
318
319 const __m256i vi10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i10));
320 const __m256i vk10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 320 * sizeof(int8_t))));
321 const __m256i vi10x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i10 + 8)));
322 const __m256i vk10x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 328 * sizeof(int8_t))));
323 const __m256i vi10xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i10 + 16)));
324 const __m256i vk10xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 336 * sizeof(int8_t))));
325 const __m256i vi10xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i10 + 24)));
326 const __m256i vk10xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 344 * sizeof(int8_t))));
327 i10 += 32;
328
329 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
330 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi10x89ABCDEF, vk10x89ABCDEF));
331 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi10xGHIJKLMN, vk10xGHIJKLMN));
332 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi10xOPQRSTUV, vk10xOPQRSTUV));
333
334 const __m256i vi11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i11));
335 const __m256i vk11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 352 * sizeof(int8_t))));
336 const __m256i vi11x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i11 + 8)));
337 const __m256i vk11x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 360 * sizeof(int8_t))));
338 const __m256i vi11xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i11 + 16)));
339 const __m256i vk11xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 368 * sizeof(int8_t))));
340 const __m256i vi11xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i11 + 24)));
341 const __m256i vk11xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 376 * sizeof(int8_t))));
342 i11 += 32;
343
344 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
345 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi11x89ABCDEF, vk11x89ABCDEF));
346 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi11xGHIJKLMN, vk11xGHIJKLMN));
347 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi11xOPQRSTUV, vk11xOPQRSTUV));
348
349 const __m256i vi12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i12));
350 const __m256i vk12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 384 * sizeof(int8_t))));
351 const __m256i vi12x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i12 + 8)));
352 const __m256i vk12x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 392 * sizeof(int8_t))));
353 const __m256i vi12xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i12 + 16)));
354 const __m256i vk12xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 400 * sizeof(int8_t))));
355 const __m256i vi12xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i12 + 24)));
356 const __m256i vk12xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 408 * sizeof(int8_t))));
357 i12 += 32;
358
359 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
360 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi12x89ABCDEF, vk12x89ABCDEF));
361 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi12xGHIJKLMN, vk12xGHIJKLMN));
362 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi12xOPQRSTUV, vk12xOPQRSTUV));
363
364 const __m256i vi13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i13));
365 const __m256i vk13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 416 * sizeof(int8_t))));
366 const __m256i vi13x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i13 + 8)));
367 const __m256i vk13x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 424 * sizeof(int8_t))));
368 const __m256i vi13xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i13 + 16)));
369 const __m256i vk13xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 432 * sizeof(int8_t))));
370 const __m256i vi13xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i13 + 24)));
371 const __m256i vk13xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 440 * sizeof(int8_t))));
372 i13 += 32;
373
374 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
375 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi13x89ABCDEF, vk13x89ABCDEF));
376 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi13xGHIJKLMN, vk13xGHIJKLMN));
377 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi13xOPQRSTUV, vk13xOPQRSTUV));
378
379 const __m256i vi14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i14));
380 const __m256i vk14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 448 * sizeof(int8_t))));
381 const __m256i vi14x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i14 + 8)));
382 const __m256i vk14x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 456 * sizeof(int8_t))));
383 const __m256i vi14xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i14 + 16)));
384 const __m256i vk14xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 464 * sizeof(int8_t))));
385 const __m256i vi14xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i14 + 24)));
386 const __m256i vk14xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 472 * sizeof(int8_t))));
387 i14 += 32;
388
389 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
390 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi14x89ABCDEF, vk14x89ABCDEF));
391 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi14xGHIJKLMN, vk14xGHIJKLMN));
392 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi14xOPQRSTUV, vk14xOPQRSTUV));
393
394 const __m256i vi15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i15));
395 const __m256i vk15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 480 * sizeof(int8_t))));
396 const __m256i vi15x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i15 + 8)));
397 const __m256i vk15x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 488 * sizeof(int8_t))));
398 const __m256i vi15xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i15 + 16)));
399 const __m256i vk15xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 496 * sizeof(int8_t))));
400 const __m256i vi15xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i15 + 24)));
401 const __m256i vk15xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 504 * sizeof(int8_t))));
402 i15 += 32;
403
404 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
405 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi15x89ABCDEF, vk15x89ABCDEF));
406 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi15xGHIJKLMN, vk15xGHIJKLMN));
407 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi15xOPQRSTUV, vk15xOPQRSTUV));
408
409 const __m256i vi16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i16));
410 const __m256i vk16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 512 * sizeof(int8_t))));
411 const __m256i vi16x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i16 + 8)));
412 const __m256i vk16x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 520 * sizeof(int8_t))));
413 const __m256i vi16xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i16 + 16)));
414 const __m256i vk16xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 528 * sizeof(int8_t))));
415 const __m256i vi16xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i16 + 24)));
416 const __m256i vk16xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 536 * sizeof(int8_t))));
417 i16 += 32;
418
419 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
420 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi16x89ABCDEF, vk16x89ABCDEF));
421 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi16xGHIJKLMN, vk16xGHIJKLMN));
422 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi16xOPQRSTUV, vk16xOPQRSTUV));
423
424 const __m256i vi17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i17));
425 const __m256i vk17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 544 * sizeof(int8_t))));
426 const __m256i vi17x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i17 + 8)));
427 const __m256i vk17x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 552 * sizeof(int8_t))));
428 const __m256i vi17xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i17 + 16)));
429 const __m256i vk17xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 560 * sizeof(int8_t))));
430 const __m256i vi17xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i17 + 24)));
431 const __m256i vk17xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 568 * sizeof(int8_t))));
432 i17 += 32;
433
434 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
435 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi17x89ABCDEF, vk17x89ABCDEF));
436 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi17xGHIJKLMN, vk17xGHIJKLMN));
437 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi17xOPQRSTUV, vk17xOPQRSTUV));
438
439 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18));
440 const __m256i vk18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 576 * sizeof(int8_t))));
441 const __m256i vi18x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i18 + 8)));
442 const __m256i vk18x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 584 * sizeof(int8_t))));
443 const __m256i vi18xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i18 + 16)));
444 const __m256i vk18xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 592 * sizeof(int8_t))));
445 const __m256i vi18xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i18 + 24)));
446 const __m256i vk18xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 600 * sizeof(int8_t))));
447 i18 += 32;
448
449 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
450 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi18x89ABCDEF, vk18x89ABCDEF));
451 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi18xGHIJKLMN, vk18xGHIJKLMN));
452 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi18xOPQRSTUV, vk18xOPQRSTUV));
453
454 const __m256i vi19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i19));
455 const __m256i vk19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 608 * sizeof(int8_t))));
456 const __m256i vi19x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i19 + 8)));
457 const __m256i vk19x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 616 * sizeof(int8_t))));
458 const __m256i vi19xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i19 + 16)));
459 const __m256i vk19xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 624 * sizeof(int8_t))));
460 const __m256i vi19xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i19 + 24)));
461 const __m256i vk19xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 632 * sizeof(int8_t))));
462 i19 += 32;
463
464 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
465 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi19x89ABCDEF, vk19x89ABCDEF));
466 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi19xGHIJKLMN, vk19xGHIJKLMN));
467 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi19xOPQRSTUV, vk19xOPQRSTUV));
468
469 const __m256i vi20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i20));
470 const __m256i vk20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 640 * sizeof(int8_t))));
471 const __m256i vi20x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i20 + 8)));
472 const __m256i vk20x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 648 * sizeof(int8_t))));
473 const __m256i vi20xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i20 + 16)));
474 const __m256i vk20xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 656 * sizeof(int8_t))));
475 const __m256i vi20xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i20 + 24)));
476 const __m256i vk20xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 664 * sizeof(int8_t))));
477 i20 += 32;
478
479 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
480 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi20x89ABCDEF, vk20x89ABCDEF));
481 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi20xGHIJKLMN, vk20xGHIJKLMN));
482 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi20xOPQRSTUV, vk20xOPQRSTUV));
483
484 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21));
485 const __m256i vk21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 672 * sizeof(int8_t))));
486 const __m256i vi21x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i21 + 8)));
487 const __m256i vk21x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 680 * sizeof(int8_t))));
488 const __m256i vi21xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i21 + 16)));
489 const __m256i vk21xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 688 * sizeof(int8_t))));
490 const __m256i vi21xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i21 + 24)));
491 const __m256i vk21xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 696 * sizeof(int8_t))));
492 i21 += 32;
493
494 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
495 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi21x89ABCDEF, vk21x89ABCDEF));
496 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi21xGHIJKLMN, vk21xGHIJKLMN));
497 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi21xOPQRSTUV, vk21xOPQRSTUV));
498
499 const __m256i vi22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i22));
500 const __m256i vk22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 704 * sizeof(int8_t))));
501 const __m256i vi22x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i22 + 8)));
502 const __m256i vk22x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 712 * sizeof(int8_t))));
503 const __m256i vi22xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i22 + 16)));
504 const __m256i vk22xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 720 * sizeof(int8_t))));
505 const __m256i vi22xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i22 + 24)));
506 const __m256i vk22xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 728 * sizeof(int8_t))));
507 i22 += 32;
508
509 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
510 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi22x89ABCDEF, vk22x89ABCDEF));
511 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi22xGHIJKLMN, vk22xGHIJKLMN));
512 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi22xOPQRSTUV, vk22xOPQRSTUV));
513
514 const __m256i vi23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i23));
515 const __m256i vk23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 736 * sizeof(int8_t))));
516 const __m256i vi23x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i23 + 8)));
517 const __m256i vk23x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 744 * sizeof(int8_t))));
518 const __m256i vi23xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i23 + 16)));
519 const __m256i vk23xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 752 * sizeof(int8_t))));
520 const __m256i vi23xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i23 + 24)));
521 const __m256i vk23xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 760 * sizeof(int8_t))));
522 i23 += 32;
523
524 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
525 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi23x89ABCDEF, vk23x89ABCDEF));
526 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi23xGHIJKLMN, vk23xGHIJKLMN));
527 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi23xOPQRSTUV, vk23xOPQRSTUV));
528
529 const __m256i vi24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i24));
530 const __m256i vk24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 768 * sizeof(int8_t))));
531 const __m256i vi24x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i24 + 8)));
532 const __m256i vk24x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 776 * sizeof(int8_t))));
533 const __m256i vi24xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i24 + 16)));
534 const __m256i vk24xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 784 * sizeof(int8_t))));
535 const __m256i vi24xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i24 + 24)));
536 const __m256i vk24xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 792 * sizeof(int8_t))));
537 i24 += 32;
538
539 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
540 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi24x89ABCDEF, vk24x89ABCDEF));
541 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi24xGHIJKLMN, vk24xGHIJKLMN));
542 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi24xOPQRSTUV, vk24xOPQRSTUV));
543
544 w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t));
545
546 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
547 __m256 vscaled89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
548 __m256 vscaledGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
549 __m256 vscaledOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
550
551 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
552 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
553 vscaled89ABCDEF = _mm256_mul_ps(vscaled89ABCDEF, vscale);
554 vscaledGHIJKLMN = _mm256_mul_ps(vscaledGHIJKLMN, vscale);
555 vscaledOPQRSTUV = _mm256_mul_ps(vscaledOPQRSTUV, vscale);
556
557 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
558 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
559 vscaled89ABCDEF = _mm256_min_ps(vscaled89ABCDEF, voutput_max_less_zero_point);
560 vscaledGHIJKLMN = _mm256_min_ps(vscaledGHIJKLMN, voutput_max_less_zero_point);
561 vscaledOPQRSTUV = _mm256_min_ps(vscaledOPQRSTUV, voutput_max_less_zero_point);
562
563 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
564 vacc89ABCDEF = _mm256_cvtps_epi32(vscaled89ABCDEF);
565 vaccGHIJKLMN = _mm256_cvtps_epi32(vscaledGHIJKLMN);
566 vaccOPQRSTUV = _mm256_cvtps_epi32(vscaledOPQRSTUV);
567
568 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
569 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
570 __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
571
572 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
573 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
574
575 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
576 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
577 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
578
579 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
580 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
581 output += 32;
582 }
583 if XNN_UNLIKELY(c != 0) {
584 const int8_t* k = (const int8_t*) ((const int32_t*) w + 32);
585 do {
586 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
587
588
589 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0));
590 const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) k));
591 i0 += 8;
592
593 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
594
595 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1));
596 const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 32)));
597 i1 += 8;
598
599 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
600
601 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2));
602 const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 64)));
603 i2 += 8;
604
605 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
606
607 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3));
608 const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 96)));
609 i3 += 8;
610
611 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
612
613 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4));
614 const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 128)));
615 i4 += 8;
616
617 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
618
619 const __m256i vi5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i5));
620 const __m256i vk5x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 160)));
621 i5 += 8;
622
623 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
624
625 const __m256i vi6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i6));
626 const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 192)));
627 i6 += 8;
628
629 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
630
631 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7));
632 const __m256i vk7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 224)));
633 i7 += 8;
634
635 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
636
637 const __m256i vi8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i8));
638 const __m256i vk8x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 256)));
639 i8 += 8;
640
641 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
642
643 const __m256i vi9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i9));
644 const __m256i vk9x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 288)));
645 i9 += 8;
646
647 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi9x01234567, vk9x01234567));
648
649 const __m256i vi10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i10));
650 const __m256i vk10x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 320)));
651 i10 += 8;
652
653 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi10x01234567, vk10x01234567));
654
655 const __m256i vi11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i11));
656 const __m256i vk11x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 352)));
657 i11 += 8;
658
659 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi11x01234567, vk11x01234567));
660
661 const __m256i vi12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i12));
662 const __m256i vk12x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 384)));
663 i12 += 8;
664
665 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi12x01234567, vk12x01234567));
666
667 const __m256i vi13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i13));
668 const __m256i vk13x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 416)));
669 i13 += 8;
670
671 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi13x01234567, vk13x01234567));
672
673 const __m256i vi14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i14));
674 const __m256i vk14x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 448)));
675 i14 += 8;
676
677 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi14x01234567, vk14x01234567));
678
679 const __m256i vi15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i15));
680 const __m256i vk15x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 480)));
681 i15 += 8;
682
683 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi15x01234567, vk15x01234567));
684
685 const __m256i vi16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i16));
686 const __m256i vk16x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 512)));
687 i16 += 8;
688
689 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi16x01234567, vk16x01234567));
690
691 const __m256i vi17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i17));
692 const __m256i vk17x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 544)));
693 i17 += 8;
694
695 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi17x01234567, vk17x01234567));
696
697 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18));
698 const __m256i vk18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 576)));
699 i18 += 8;
700
701 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi18x01234567, vk18x01234567));
702
703 const __m256i vi19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i19));
704 const __m256i vk19x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 608)));
705 i19 += 8;
706
707 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi19x01234567, vk19x01234567));
708
709 const __m256i vi20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i20));
710 const __m256i vk20x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 640)));
711 i20 += 8;
712
713 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi20x01234567, vk20x01234567));
714
715 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21));
716 const __m256i vk21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 672)));
717 i21 += 8;
718
719 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi21x01234567, vk21x01234567));
720
721 const __m256i vi22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i22));
722 const __m256i vk22x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 704)));
723 i22 += 8;
724
725 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi22x01234567, vk22x01234567));
726
727 const __m256i vi23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i23));
728 const __m256i vk23x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 736)));
729 i23 += 8;
730
731 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi23x01234567, vk23x01234567));
732
733 const __m256i vi24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i24));
734 const __m256i vk24x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + 768)));
735 i24 += 8;
736
737 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi24x01234567, vk24x01234567));
738
739 k += 8;
740
741 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
742 vscaled01234567 = _mm256_mul_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.scale));
743 vscaled01234567 = _mm256_min_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point));
744 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
745
746 w = (const void*) ((const int32_t*) w + 8);
747
748 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
749 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
750
751 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
752
753 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
754 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
755
756 if XNN_LIKELY(c >= 8) {
757 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
758 output += 8;
759 c -= 8;
760 } else {
761 if (c & 4) {
762 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
763 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
764 output += 4;
765 }
766 if (c & 2) {
767 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
768 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
769 output += 2;
770 }
771 if (c & 1) {
772 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
773 output += 1;
774 }
775 c = 0;
776 }
777 } while (c != 0);
778 }
779
780 output = (int8_t*) ((uintptr_t) output + output_increment);
781 } while (--output_width != 0);
782 }
783